[ProfSynapse]

Change

Resolve the effective GRPOConfig precision before loading the model. When the retained model dtype is null, pass float16 or bfloat16 to Unsloth's loader to match the trainer. Preserve an explicitly configured model dtype. Refresh only the GRPO offline/outer closure hashes and sync the fine-tuning skill mirrors.

Evidence and limits

The consumed 0d515 attempt ended with closed GRPO_TRAINER_CHILD_EXEC_TYPE_TRAINER_SETUP, retry false, no artifacts or parameter/reward verdict. It does not retain the exception message or observed loaded dtype, so the historical cause remains unproven.

Pinned published wheels inspected offline: TRL 0.24.0, Transformers 4.57.1, Unsloth 2026.1.2. The exact Unsloth generated constructor guard, extracted from the pinned wheel and executed without ML imports, raises TypeError for a BF16 model with fp16 trainer flags and accepts matching float16/fp16 and bfloat16/bf16 pairs. The pinned Unsloth loader auto-selects BF16 on BF16-capable hardware when dtype is null; the retained workload asks fp16. The pinned Transformers loader updates config.dtype when an explicit dtype is passed. This identifies a reproducible constructor incompatibility, not a retrospective proof of which historical TypeError fired.

Verification

• 70 focused provider-free tests passed; one optional ML-dependent module skipped in the CPU environment.
• GRPO offline and outer closure maintainers CURRENT; 131 and 13 members retain order and metadata. SFT, HF parent, Modal training runtime and inference lock maintainers CURRENT with no changes.
• Canonical skills synced and checked; git diff --check clean.

No GPU call, provider action, image build/pull, model call, package upgrade, training replay or old-resource change. A fresh qualification requires a new clean host gitlink/source commit, reviewed release/image/profile and independent admission, then the usual bounded CPU and one-shot GPU gates. Keep the consumed attempt's claims, resources and read keys.